Frontier Data Efforts

Data Markets are entrenched. Every company will train their own models. A new class of problems exist with the normalization of RL. I'm building something new to tackle:

1. Evals aren't aligned properly to real world tasking
2. Contributor incentives are misaligned and mostly viewed as human problems rather than technical problems
3. Reward Rubrics are increasingly hard to define for RL env companies because they aren't domain experts. They need to be defined by enterprise domain experts who are actually at the frontier.
4. The shape of good data is subjective and hard to standardize. Why don't we have evals and benchmarks for good data before they get submitted to NeurIps?

I'm working on frontier solutions to curate real world datasets and benchmarks to solve these problems. These happen to do with issues regarding enterprise interactions with frontier RL problems - reward model alignment with business KPIs, maintainable envs, and evals for "data realism" as it pertains to real economic value.

Reach out to cr4sean@gmail.com with proposals/thoughts. Much of our work is based on my observations in human data markets from spending time with top RL env companies and human data pioneers.